Visual-inertial SLAM is a common way to give drones accurate position estimates without expensive sensors, but most existing systems rely on cameras with a narrow field of view. That makes them prone to drift or outright failure when a UAV flies through changing light, sharp maneuvers, or long trajectories. The authors of a new arXiv preprint argue that panoramic cameras, which capture the full surroundings, are a natural fix—yet panoramic VI-SLAM for drones has been little explored.

To close that gap, the team built PanoAir, a panoramic VI-SLAM framework that uses omnidirectional images for both feature extraction and loop closure. They also collected a real-world UAV dataset spanning varied illumination, altitudes, trajectory lengths, and motion dynamics. In experiments, PanoAir outperformed existing baselines on both their dataset and public benchmarks, completing all 17 sequences without a single tracking failure and cutting absolute trajectory error by up to 47.1% compared with the best-performing baseline.

The authors also tested PanoAir on an embedded platform, showing it runs with computational efficiency comparable to a desktop PC, which suggests it could be practical for onboard drone use. The source code and dataset are publicly available, giving other researchers a foundation for further work on panoramic visual-inertial navigation.